Convolutional layers by themselves are nothing more than linear classifiers. 1 x 1 convolutions allow one to introduce non linear classification between convolutional layers. They act as a fully connected layer. This allows the data seperation boundaries to be “curved” towards a more perfect fit.
A fully connected layer is a layer where every node, is connected to every node in the next layer. Fully connected layers are non-linear classifiers that allow one to draw highly complex seperation lines between multiple types of data. Fully connected layers tend to do best when trained on labeled data; I.E learning that is fully supervised. This is becuase multilayered perceptrons depend heavily of back propegation; the calculation of the partial derivative distance between the layer prediction and the actual label.
I believe the classifier would struggle in such a situation. While its layers may contain classifiers that allow it rule out “crowd” objects. I would lack the ability to identify the actual target due to the fact that it would be unable to identify it’s feature.
This is the my submission repository for my Robot arm pick and place project.
The purpose of the project was to collect data for a train a fully-convolutional network that would enable a drone to identify a particular target “person” and track them.
For the data collection part of this project, I created a patrol circuit around the enviroment’s parkinglot area. The circuit consisted three loops, one at ground level, one at twice “person hieght” abouts and one at simulation starting level. Inside the patrol circiut I created a zigzagging hero patrol path and a number of spawn pionts interspersed around the hero path. I then recorded the simulation for three hours obtaining a collection of 4100 images. This I deposited in run 1.
My technique for collecting data, had some strengths and weaknesses. For one, the ground level images might have actually degraded the accuracy of my model, because the test drone always remains at the same height. Furthermore a more, concentrated collection area would have improved my model’s ability to identify the target in a crowd. That said, my data set contains a large number of images of the target at a distance; which may have improved the model’s ability to identify targets at a distance. Hopefully, have 4100 extra training images would improve the model’s performance overall.
Collection Example 1